Feat/autoprompt classifications on conversations

[Omer-ler]

Goal: run an auto-prompt procedure on conversations (instead of a single simple responses).
also: a new feature - combining OAI's structured outputs in the auto-prompt pipeline.

TODO:

1. Code cleanup
2. get rid of the hard coded "5" limitation (in two places)
3. get rid of prints

usage:

export OPENAI_API_KEY=...
python main.py --prompt "Does the User request to finish the game? answer 'yes' or 'no'" --task_description "Assistant is an expert classifier that will classify if the user, a 5 years old child, want to finish the game or not." --output_path finish_request_rlaif_opposite_labels --basic_config_path config/config_auto_sync_small.yml --rlaif_instructions "Assess whether the User would like to finish the interaction. Answer 'yes' if it does and 'no' otherwise."

Example run:
wandb: 🚀 View run devoted-music-60 at: https://wandb.ai/omer-ler/AutoGPT/runs/zqfsmdfx